GPUs & CPUs & Enthusiast hardware: Questions, Discussion and fanboy slap-fights - Nvidia & AMD & Intel - Separe but Equal. Intel rides in the back of the bus.

[The Mass Shooter Ron Soye]

the day you can get a CPU or APU with the performance of a 3060 is the day the GPU market starts dying.

Right now the best right now is the AMD Ryzen™ 7 8700G which has an iGPU called the 780M which is pretty close to a RTX 2050 mobile in power

AMD’s AI Max+ 395 has the performance of around a 4060-4070. Cost’s over $1k.

It's not a "mainstream APU" but it would have an effect on low-end GPUs if it was low cost, which it clearly isn't to start. It can't go on the AM5 socket (which doesn't support quad-channel anyway) but it could be soldered onto a motherboard like those Minisforum PCs/boards that use the 7945HX.

Strix Halo uses up to 2x modified Zen 5 CCDs at around 67mm^2 each, and the big I/O die is around 308mm^2. They are probably using N4X for the CCDs and N3E(?) for the IOD. It will have higher packaging costs from using silicon bridges in the base die, but this is probably identical to InFO-RDL/"Infinity Links" which was used in RDNA3 cards. A non-TSMC company may be providing this.

That I/O die size is on par with a Navi 31 GCD, e.g. the 7900 XTX. Lower cost Strix Halo can ditch a CCD, although you lose 20% of the CUs (32/40) due to segmentation. Pair it with the bare minimum of 32 GB LPDDR5X (it would be better if 48 GB was an option), and you may have a complete system under $1k... a year from now after demand settles.

But this is AMD's first attempt at a PC mega APU. Meanwhile, the smaller Strix Point is faster than the 8700G and you could also get that in a mini PC. Next-gen Medusa Point is leaked to use a 12-core desktop CCD, ditching the big/small cores in split CCXs, and may finally add some Infinity Cache to a mainstream APU (not confirmed).

 

[The Ugly One]

Just tell them your use case is AI, so they’ll just have to replace it with an H100 if they can’t find you another 4090.

Up to an 18 month wait on those bad boys. But maybe the AI bubble will pop first. I keep hearing in places that matter, "Okay, everyone's spent billions on AI infrastructure. Now we just need to find a way to show them how to make money with it." This is bubble behavior. Everyone's apparently been buying GPUs and software licenses out of pure FOMO, with no idea how any of this stuff will actually benefit their business.

I know of one company that bought a DGX and is paying for ChatGPT to generate marketing material, and it's costing them even more money, since the output is such crap that they have to have professionals review and edit everything anyway.

 

[Gerelatha]

Incidentally, the review card batches were NOT affected by the Nvidia ROP problem.

https://www.tomshardware.com/pc-components/gpus/some-rtx-5090s-are-shipping-with-missing-rop-units-leading-to-less-gaming-performance-report

If they intentionally cut down consumer chips while giving reviewers better chips there would be a class action lawsuit, a batch of chips being defective is far more likely.

 

[jeff7989]

How would it pass QC? Shouldn't it be detected during binning? To me it seems like nvidia was seeing if they can get away with it, because they need best dies for AI.

 

[The Ugly One]

How would it pass QC? Shouldn't it be detected during binning? To me it seems like nvidia was seeing if they can get away with it, because they need best dies for AI.

"We're printing money, we don't need to do QA/QC." -at least two of my former employers

 

[The Mass Shooter Ron Soye]

Note: DLSS4 upscaling (not frame gen or multi-frame gen) works on all RTX GPUs. TL;DW: It's good. Aspects of DLSS4 Performance are sometimes better than DLSS3 Quality. Performance at the same DLSS resolution is slightly less, but DLSS4 Perf is going to outperform DLSS3 Qual. @The Ugly One I'll be interested to hear your impressions of DLSS4 upscaling after you try it out.

Strix Halo at high wattage good for 1440p, sometimes 4K.

 

[Get the rope Macaulay!]

and it's costing them even more money, since the output is such crap that they have to have professionals review and edit everything anyway.

why didn't they have those professionals in the first place? like thats what chatGPT is replacing, the useless underlings. instead of a snotty graphic designer who vapes and goes "i'll get it to ya after my lunch break" you get the same shitty quality instantaneously.

works on all RTX GPUs

meaning 2000s and above, the only "unique" shit on the 50 series is the multi-frame gen. a 2060 can be had for under $200 now.

 

[The Mass Shooter Ron Soye]

meaning 2000s and above, the only "unique" shit on the 50 series is the multi-frame gen. a 2060 can be had for under $200 now.

That's what I said, yes.

At 25:23 Hardware Unboxed man says "Based on what I saw testing ray reconstruction, I expect the new DLSS4 transformer model to perform a bit worse on older GeForce 30 series cards, but we'll have to investigate that in a future video."

 

[Fcret]

I know of one company that bought a DGX and is paying for ChatGPT to generate marketing material, and it's costing them even more money, since the output is such crap that they have to have professionals review and edit everything anyway.

This has been my problem with AI generated code. I can get it to vomit something up but in the time it takes me to validate it and make corrections, I could have just written it from scratch.

I've seen it absolutely wreck the independent thinking skills of junior devs though, so at least I'll have job security.

 

[Mister Fister]

AI can generate work, it doesn't know if it's good or how to use it.

 

[Betonhaus]

Bruh

 

[The Ugly One]

View attachment 7016711
Bruh

This is like saying a PowerMac G3 runs OS 8 faster than an M3 MacBook Pro.

 

[Mister Fister]

This is like saying a PowerMac G3 runs OS 8 faster than an M3 MacBook Pro.

I don't know who reddit hates more Jensen Huang or Donald Trump.

 

[NoBueno]

Do you think I could put fish in there?

 

[⠠⠠⠅⠑⠋⠋⠁⠇⠎ ⠠⠠⠊⠎ ⠠⠠⠁ ⠠⠠⠋⠁⠛]

This is like saying a PowerMac G3 runs OS 8 faster than an M3 MacBook Pro.

But the M3 would run MacOS 8 faster. Because it's not a step backwards

 

[Betonhaus]

But the M3 would run MacOS 8 faster. Because it's not a step backwards

technically no as there would be more latency. The latency would be largely invisible to the user, but it's still there.

 

[The Mass Shooter Ron Soye]

After the 5090 and 5070 Ti, it's the 5080's turn:
First GeForce RTX 5080 discovered with 104 ROPS, 8 missing in action (archive)

 

[4090 Chad]

After the 5090 and 5070 Ti, it's the 5080's turn:
First GeForce RTX 5080 discovered with 104 ROPS, 8 missing in action (archive)

Everyday there is new news that makes me feel even better about buying my 4090 when I did

 

[DavidS877]

After the 5090 and 5070 Ti, it's the 5080's turn:
First GeForce RTX 5080 discovered with 104 ROPS, 8 missing in action (archive)

Glad I can't find one in stock at MSRP.

 

[Just Some Other Guy]

After the 5090 and 5070 Ti, it's the 5080's turn:
First GeForce RTX 5080 discovered with 104 ROPS, 8 missing in action (archive)

The shitshow must go on! What's funny is the official Nvidia response didn't mention 5080s, so either they're lying or they actually have no clue about the full extent of the issue.